Learning Hybrid Relationships for Person Re-identification
Authors: Shuang Liu, Wenmin Huang, Zhong Zhang2172-2179
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Market-1501, Duke MTMCre ID, CUHK03 and MSMT17 demonstrate that the proposed HRNet outperforms the state-of-the-art methods. |
| Researcher Affiliation | Academia | Tianjin Key Laboratory of Wireless Mobile Communications and Power Transmission, Tianjin Normal University, Tianjin, China. shuangliu.tjnu@gmail.com, huangwenmin2018@gmail.com, zhong.zhang8848@gmail.com |
| Pseudocode | No | The paper describes the approach and algorithms using mathematical equations and textual explanations, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statements about releasing source code or a link to a code repository for their described methodology. |
| Open Datasets | Yes | Market-1501 (Zheng et al. 2015) is shot by six disjoint cameras at the Tsinghua University campus, and it consists of 32,668 images of 1,501 identities. According to the database setting, the training set contains 12,936 images of 751 identities while the test set comprises of 3,368 query images and 16,364 gallery images from the other 750 identities. Duke MTMC-re ID (Ristani et al. 2016) consists of 36,411 images of 1,404 identities, among which 16,522 images of 702 identities are utilized as the training set. And 19,889 images of 702 non-overlapping identities are treated as the test set with 2,228 query images as well as 17,661 gallery images. Furthermore, Duke MTMC-re ID is collected by eight high-resolution cameras. CUHK03 (Zhao, Ouyang, and Wang 2014) is composed of 14,097 images of 1,467 identities, and each identity is captured by two of ten cameras at the CUHK campus. According to the database setting, the training set consists of 7,365 images of 767 identities and the test set includes 1,400 query images and 5,332 gallery images of 700 identities. CUHK03 provides two types of annotations for all images, i.e., manually labeled bounding-boxes and DPM-detected bounding-boxes. In this work, we evaluate the proposed HRNet on DPM-detected bounding-boxes which are more challenging. MSMT17 (Wei et al. 2018) comprehends 126,441 images of 4,101 identities from 15 cameras and it is divided into the training set including 32,621 images of 1,041 identities and the test set including 93,820 images of 3,060 identities. |
| Dataset Splits | No | The paper explicitly states the training and test set sizes for each database. While it mentions 'validation' in the context of verification loss, it does not explicitly define a separate validation dataset split. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., specific GPU models, CPU types, or memory). |
| Software Dependencies | No | The paper mentions using 'Adam as the optimizer' and 'Res Net-50' as the CNN model. However, it does not specify any software libraries or frameworks with version numbers (e.g., PyTorch 1.9, TensorFlow 2.x). |
| Experiment Setup | Yes | The batch size is set to 66 where we randomly select 11 identities and 6 images for each identity, and the epoch number is set to 200. We adopt the random cropping and the horizontal flipping for data augmentation. We utilize Adam as the optimizer and set the weight decay to 5 10 4. The learning rate is initialized to 3.5 10 4 and it is decreased by the factor of 0.1 at the 40-th and 120-th epochs. The pre-defined margins ω in Eq. 4 and θ in Eq. 12 are both set to 0.3, and the hyperparameters α and β in Eq. 7 are set to 0.9 and 0.8, respectively. |